M12.3: persist exact population state and automatic conversion policy
Foundation / checks (deploy) (push) Successful in 5s
Foundation / checks (ecs) (push) Successful in 3m16s
Foundation / postgres-tools (push) Canceled after 0s
Foundation / server-image (push) Canceled after 0s
Foundation / image (push) Canceled after 0s
Foundation / checks (client) (push) Canceled after 13m8s
Foundation / checks (checks) (push) Canceled after 13m12s
Foundation / checks (postgres) (push) Canceled after 13m2s
Foundation / checks (network) (push) Canceled after 13m4s
Foundation / checks (tsan) (push) Canceled after 13m9s

This commit is contained in:
dingo
2026-10-17 12:11:10 -04:00
parent 2b074b9ae1
commit 5decaa0326
14 changed files with 624 additions and 15 deletions
+11
View File
@@ -1,5 +1,16 @@
# Project status and next steps
## M12.3a exact conversion recovery - 2026-10-17
Population envelope 1 retains biology, cohort membership, policy and conversion
sequence together. Detached candidates preserve their predecessor; opt-in
automatic hysteresis is deterministic and disabled by default. Separate-process
continuation matches uninterrupted execution, with malformed input, oscillation
and sequence exhaustion checks. Release tests pass 74/74, plus clang-format/tidy
19 and ASan/UBSan. [ADR-0140](adr/0140-exact-population-checkpoint-and-policy.md)
records the remaining production profile/trace/publication adapter; local file
recovery does not close that gate. No deployment or live format change.
## M12.2d exact cellular cohorts - 2026-10-17
Under the maintainer's delegated M12.2 authority, ADR-0139 selects exact local
@@ -0,0 +1,47 @@
# ADR-0140: Exact population checkpoint and automatic local policy
## Status
Accepted for the portable M12.3 implementation, 2026-10-17. Production publication
and enabling remain separate gates. No live region changes or deployment.
## Decision
Persist exact biology and exclusive cohort membership together in population
envelope 1, never in a sidecar that can get ahead of its biological checkpoint.
Retain the existing nested rules-8 owner bytes without reinterpretation. Store
policy configuration and a monotonic conversion sequence. Recovery restores
the exact committed membership sample; it does not repeat a random allocation.
Exact members have no fractional demographic residual or sampling stream.
Provide detached conversion candidates: the predecessor remains untouched until
the caller publishes the entire candidate under the current durable ownership
fence. Pre-publication loss recovers the predecessor; confirmed publication must
recover the complete successor. In-memory success is not a durable receipt.
Sequence exhaustion refuses conversion before mutation. Unknown versions,
malformed bounds, noncanonical member IDs, wrong identity/tick, ineligible
members and corrupt checksums are refused.
An opt-in automatic policy runs only after a successful biological tick. Whole
wildlife genome groups materialize if any member is within 1200 micrometres of
an active/lingering controlled observer. They collapse if all members are beyond
1800 micrometres of every such observer, including when none exist. The band
retains current membership. Distances use lower-left positions in the existing
local coordinate frame. Suspended controlled lives are not observers and never
become cohort members. Unsupported wildlife presence keeps the group individual.
Radii are configurable, bounded, persisted trial policy rather than locked
biological constants. No state, phenotype, resource or ancestry changes at a
conversion. Defaults keep automation disabled.
## Boundary
The portable owner, codec and separate-process continuation checks implement
state closure. They do not supply a new PostgreSQL publication profile, durable
conversion trace, database fencing, lost-reply reconciliation or host integration.
Do not place this envelope under existing profile 4: that profile contains raw
CellularLives bytes. The new payload needs a dedicated profile/schema migration
and recovery/publication adapter before production use. Existing M7/M11 stores
and live sessions remain on their existing accepted formats.
This explicit boundary permits engineering work on the exact representation
without mistaking a local file test for production crash-consistent durability.
+2
View File
@@ -157,3 +157,5 @@ design. No unadopted candidate is a link dependency of the proof.
- [ADR-0085: Rebindable engineering controls](0085-rebindable-engineering-controls.md) - SDL-free action/HID-key bindings, conflict validation, versioned all-or-nothing controls file, binding capture before toolkit/world.
- [ADR-0139: Exact cellular local cohorts](0139-exact-cellular-local-cohorts.md)
- [ADR-0140: Exact population checkpoint and policy](0140-exact-population-checkpoint-and-policy.md)
+17 -3
View File
@@ -20,9 +20,23 @@ removes stale ownership tags; funded births begin as individuals. Suspension
continues to freeze controlled biology regardless of representation elsewhere.
The model retains exact members and makes no aggregation speedup claim. There is
no regional population approximation, process migration, production selector or
new persisted format in this segment. The existing 256-life ceiling remains;
target-host service capacity is a separate qualification.
no regional population approximation, process migration or production selector.
The existing 256-life ceiling remains; target-host service capacity is separate.
M12.3 adds [population envelope 1](../../schemas/cellular-population.md), detached
conversion candidates and an opt-in automatic policy (ADR-0140). The default
remains disabled. The policy materializes whole genome groups within 1200 um of
any nonsuspended controlled observer, collapses beyond 1800 um of all observers,
and retains membership in the intervening band. Biology always advances through
the same 50 ms kernel. Policy configuration, sequence, membership and complete
biology restore together; loading a checkpoint does not reselect members.
`conversion_candidate` leaves the predecessor intact and returns the complete
exact successor for fenced publication. Successful candidate creation and local
checkpoint file writes are not durable production acknowledgements. A dedicated
PostgreSQL profile/trace/publication adapter is still required; existing raw
cellular profile 4 cannot contain this envelope. The separate-process test
demonstrates identical continuation, not power-loss or database-failover safety.
Run `evo-cellular-cohorts --verify` for the CTest matrix or omit the option for
8192 ticks per case. The executable compares complete canonical state every tick,
+1 -1
View File
@@ -10,7 +10,7 @@ Commit and push each validated scoped segment separately.
|---|---|---|
| 12.1 | Engineering evidence policy and explicit deferrals documented | ADR-0133, identity/time contract and worked histories; no new game rules or runtime versions. |
| 12.2 | Exact local model selected and paired qualification implemented (12.2d); production cost gate open | ADR-0139, exact canonical equality across 196,608 paired ticks per platform. Regional approximation disabled; no compression claim. Deployment-host capacity and fuller interaction coverage remain open. |
| 12.3 | Exact-owner durable implementation authorized; production enabling depends on qualification | ADR-0139 resolves the averaging conflict for retained members. Durable conversion closure/publication remains to implement. |
| 12.3 | Portable exact conversion closure, automatic policy and process recovery implemented | ADR-0140, population envelope 1. Dedicated PostgreSQL profile/trace/publication adapter and fault injection remain required before enabling the live selector. |
| 12.4 | Blocked on 12.2-12.3 | Existing cellular worker pauses without sessions/lingering players; bounded unattended recovery and biological migration escrow remain unimplemented. |
| 12.5 | Blocked on 12.2-12.4 | Finite M6 process fixtures do not qualify coupled persistent cellular/cohort atmosphere/water budgets. |
+24
View File
@@ -1,5 +1,29 @@
# Validation report
## 2026-10-17 - M12.3a exact population checkpoint closure
ADR-0140 adds population envelope 1, detached conversion candidates and opt-in
automatic hysteresis to portable lifecycle orchestration. Existing raw cellular
profile 4, database schemas, biology, goldens and live configuration are unchanged.
Executed: MSVC 19.44 Release configure/build and **74/74 CTests**; GCC 14.2
Release build and cohort/transition/process-restart CTests; clang-tidy 19 analysis
build, repository clang-format 19 and the two new CTests (3.29 s); ASan/UBSan
transition/restart CTests (12.10 s), no reports. Explicit checks work under
NDEBUG. Codec tests include sampled truncations, appended bytes, corrupt checksum,
invalid version/flags/policy/identity/tick/length/count/member, controlled-member
forgery and sequence exhaustion. One hundred conversion/restore oscillations
leave all biological bytes unchanged. Automatic policy matches the active
biological control across 513 offline ticks and 32 checkpoint restores.
The fresh-process test writes a cohort-bearing checkpoint after 250 offline
ticks, exits, restores in a new executable process and advances 128 more ticks;
the entire resulting population envelope equals uninterrupted continuation.
It tests state closure, not fsync/power loss or a database publication fence.
There is no new PostgreSQL profile/trace/publication adapter yet. Production
crash-boundary, stale-writer and lost-reply acceptance remain open M12.3 work.
Contract: [population envelope](../../schemas/cellular-population.md).
## 2026-10-17 - M12.2d exact cellular cohorts
ADR-0139 selects exact retained-member local cohorts under delegated M12.2